Infrared Small Target Detection, High-Precision Localization and Segmentation: Using TDU Kernel

Abstract

Aiming at the challenge of infrared small target detection with different shape and size under the different scene, a novel algorithm architecture is proposed using the kernel of Target Detection Unit (TDU). The TDU incorporates the fractal geometry design and dual-scale structure, which can execute three main sub-tasks: preliminary target detection, target localization with high precision and target segmentation by pixel-level. First, the principle establishes a dual-scale target detection structure, selects the central point, decomposes the scale information and constructs the Integrated Local Contrast Saliency (ILCS) map, the target preliminary result is obtained by the visual attention mechanism of “top to bottom”. Second, the principle adopts the scale-recursion algorithm by the mechanism of “bottom to up” to locate the target precisely from the preliminary result along with Area Optimal Recommend Mechanism (AORM) strategy. At last, the separated local histogram is used to segment the target by per-pixel with suitable threshold. From the experimental results, conducted across five different types of infrared-scenes including infrared sky scene, infrared maritime scene, backlight illuminance scene, infrared scene with interference and infrared scene with small and dim target, we observe the performance of high accuracy rate and remarkable robustness.